 Now we start. I want to welcome today the BioXcel webinar number 43. The presenter is Christian Blau and he will be speaking about density guide simulation, combined in cryo-EMDATH and molecular dynamics simulation. This webinar is a part of a series of webinars brought up by BioXcel, that is the European Centre of Excellence for Computational and Biomolecular Research. And with this series of seminars, we can try to cover different topics of computational biomolecular simulation. And so everything that you think might be interesting to listen in this webinar, you are welcome to post on our chat or ask BioXcel or our forum or just send an email. So I hope that the seminar will be of your interest. So now we start. I would like to start with introducing Christian. Christian belongs to the Gromax team and he has done his PhD in Gottingen in a group of Grutmuller in the Max Planck Institute for Biophysical Chemistry. And since 2015 he moved to Sweden and working in the group of Erik Lindel, in particular in the first Stockholm University at the moment and KTH. His main interest is the development of the Gromax software package and in particular as a particular interest in understanding user driven development and also in understanding how simulation and experiment can be combined. And this is probably what he will deal in this seminar. Please Christian, now I will provide Christian. I will give the Christian the option to start. I will share. Thank you so much for the nice introduction and hello everybody. I hope you're all doing well and all doing fine. We will start in a second. So in this talk it will be very practical. So this will be a very detailed practical talk on how to actually do things for the webinar. So you see also some things. So right now I will really go into how step by step and practically you can set up a simulation so that it is run with a certain given aquarium density and you might use this for all sorts of densities. I might be talking about aquarium densities because this is our main working course but you can use whatever type of density data. Some things you might want to do here is that you might try to automate model building of structures against the aquarium maps especially when you have a similar structures already built and then a new map coming up. This might be a tool that might help you in the process quite a bit. Then another aspect of this is of course when you're model building aquarium maps you can only build exactly one model but you might be interested in all the different types of small configuration changes. The kind of thermal motion that is behind such a map that is hidden. If you only look at exactly one density and exactly one map so you might want to use the tool for this. Here we can balance between map, influence and surcameral parameters so we have a wide range with this tool on hand to explore what happens if we get more weight to the experimental data what happens if we get more weight to biochemistry considerations that are reflected in the force field. If you start looking at this whole complex of density guided simulations in Gromix the best thing to do is to read the manual. I can only recommend strongly doing that. You find two pages of the manual that are important for understanding the density guided simulations. One that is more focused on the theory and the reference manual where we really lay out the equations that devise to the forces and simulations and the other part is the more practical side of things in how to set the parameters for the MDP options. So in case I'm talking too slowly or you want to read ahead this is the most essential information to take from this webinar that is where to find the information. On hand in quite a condense manner however. What we can do now, and this is just one example of simulation where you see you have some type of structure and within the simulation right time you want to fit it into a certain density. So this is one concrete simulation result. You just see it rendered here and one other result you might see here is where you want to keep in a sum of structures close to a certain density also nothing that can be done. And I'll just guide you step by step now through the process on how to set up exactly such a type of simulation. So how does it work? How do we find models against densities? How do we calculate the forces that drive our simulations and are added to the simulations? When we only have an atom represented here in green and a reference density which are kind of different objects reside in different spaces so to say. The trick here is that we calculate a simulated density from our atoms by spreading them with gaussians as you see here in green. And once we have a simulated and a reference density we have things we can compare on equal terms where we get some type of similarity gradient measure that tells us how different into each point in space are simulated in reference density and also what we can calculate here is how does the difference change? How strong of a change in density would be required to make these two densities more similar to one another? And this is what we use to calculate the forces then that act on our atoms and you see here that the green atom then would finally receive some type of force pushing it closer to the reference density. Then one more thing to consider here so that things are not continuous with the densities but we have some kind of voxel spacing so things happen in a discrete space but the same mouth, the same ideas still hold we have to just take care to acknowledge that our integrals are approximative here. So this is the picture in 1D in multiple dimensions this is what it looks like this kind of maybe slightly mysterious picture from the beginning where you see a structure in wall and stick representation and kind of a green isosurface now in three dimensions it's hard to show this like a true density character here so we see green isosurface and we calculated from this reference structure and we want to push this structure into the gray reference density here so we compare green and gray and have some type of similarity gradient measure density I don't go into the details here you can read that up in the manual shown in yellowish so the same picture as you've seen just before in 1D now in multiple dimensions and from this we can calculate some type of gradient vector field that allows us to evaluate the forces at each and every atom so this is the nitty gritty details of how things work technically but they might help you to understand the parameters you might want to set during these simulations and in contrast to the complex theory I just showed here the setting up is really easy the base idea here is to just add density to a normal simulation setup and this is really what every simulation setup you wish to have and also the setup for these density guided simulations in this case here are done in the mdp file with all the options just called density guided simulation something so this is exactly where you want to look for information and if you really just want to go ahead just try the most minimal setup and that is just setting density guided simulation to active then all other values can have some kind of default values for example a reference density where we expect reference density in your mdron directory now the type of reference density data is very much influenced by crye em file formats xray cross walker view file formats and what I followed here is the emdb map distribution format description so you see a reference down here but what it means in practice is that for the largest part mrcccp4.map what every format that have share all the same kind of family of a density file format should work you have a format that you don't know exactly if it works or doesn't you will file drastically we check the head at the moment we read in the map format so just go ahead try your map extension if it is close to one of these standard formats then that's all fine we don't use all too much information in the density format header so that's why we are quite compatible with a bunch of maps however there are some things to be there and this comes with the whole issue of density formats evolving over a long period of time in slightly different fields also moving from x-ray crystallography to cryo electron microscopy so we have different conventions in some map formats and tools regarding translation, vectors rotation so be careful if you use a third party density editing tools clipping off things some tools that range density values if you want to make sure that your density is really treated the way it looks like in a tool I could recommend just visualizing in BMD because we use the same density data interpretation routines as VMD does and also look at your simulations obviously so if something really weird happens then it might be a density editing issue and this might be quite tricky and it's a bit subtle because there's lots of tools out there that do very different things to the densities there's yet another thing that might make your mind go burst sometimes and that is a periodic boundary conditions in gromics we like to use periodic boundary conditions to simulate the infinite systems and get rid of finite size effects but very naturally densities have no periodic boundary conditions and the question is now how do we reconcile these two so you see in lack on the right hand side of the screen some typical periodic boundary conditions set up as it happens in gromics where you have a lengthy protein in something like this very cool shape just really sketching out some PBC behavior of a system in this case in two dimensions and we have a blue reference density and what we decided to do here is to take the center of the reference density as the overall reference and then look at all atoms in your simulation system that are closest to this reference density point the blue point here and plug these into account so if you run the simulation against this kind of blue density that is just larger than your simulation box which is okay and we can treat that be a rare however that the atoms that will receive the density in forces that are part of this black boundary box here and that it means that we will see some atoms for example in the lengthy shape that will receive no forces because they are quite far away from this blue blur here so it's a periodic boundary effects here to be taken care of and to think of the easiest way out here is of course if you use roughly similarly sized maps and periodic boundary boxes for your simulations it's also an easy way to gain a bit more performance if you for example make maps slightly smaller the parameters to set they to now get started with the density simulations if you want a brief and fast overview of all the parameters that there are you can just set the density guided simulation active to true one grand pp and then have a look at what default parameters are generated for you and you see something that's reflected here on the right hand side where you see a brief explanation of the parameter and parameter value that are set and now I'll go through these parameters one by one and explain a bit what the effects are, why you want to set them what type of ranges you would like to use for certain of these parameters the first parameter to set is what part of the simulation system should be considered for the density guided simulations in most cases this might be protein but I'm sure there's a bunch of cases where these groups might be different, for example if you have a density that is only partially describing your system you can choose to also have just a portion of your system react to the density and then the rest of your system will just act as if there's no density or you might even go further and say maybe you have some type of membrane micelle around some ion channel maybe and you want to take that into account you can do that as well here by setting the density guided simulation group, by default it's protein if you go and want to simulate some material science things which you can perfectly do maybe not even thinking of crime densities of course you would want to be careful to set this group to something more reasonable as it's a default protein which might not be even part of your system there then comes one of the largest impacts on your decision on how your simulation will behave as a setting and that is how do you even compare a simulated density to a reference density and there's three different options available here you could think of many more and you could even quite easily implement more into Gromix if you're keen to do that the choices you have right now with Gromix 2020 is a some type of inner product that really makes sure that we force things into the highest density regions often might not be the very best behavior you want of the code because then you really will cramp things into much into the densities but it can be the thing you want especially if you have lots of unwanted parts of densities it has some type of snappy behavior for a better lack of a word that means that things stick to very high density regions and it's hard to move things out of the density once a helix for example has fit in there which might make more bit successful for biochemistry deformations then on the other end of the spectrum there's the relative entropy similarity that has a measure that really looks at the difference between simulated and reference density in terms of overall information entropy and that sets large focus on overall fit of the two densities and even larger focus on things regions that don't fit so that has very nice properties in terms of smoothness, gentle fittings so to say but it really tries hard to make all regions fit and fill the void so to say all types of extra density you have so if you have a very noisy density with lots of extra density that might not be so interesting for your system and not even have a counterpart to representation in your MD simulation this might cause you some trouble otherwise it's a wonderful nice way of fitting densities and as a compromise between the two and also the classical option that lots of people have used in other implementations all types of density guided simulations is a cross correlation potential that shows perhaps some properties of both the next important contributed to how your simulation is running is how you pick the spreading weights and the thinking of cryeum there's two obvious choices here unity would mean that all items are received exactly the same way so hydrogen and a hydrogen atom and a phosphor atom would all be spread out and have the same contribution to the density here this might be a useful thing especially in combination but the density fitting group if you use a protein without hydrogens so protein minus H then that gives you a good approximation to the simulated density and at some speed up benefits even because so then you have to calculate less spreading of atoms if you exclude the hydrogen atoms so that might be a useful thing to have another alternative to that is using the mass of the atoms to roughly and have a rough approximation to kind of a scattering properties scattering factors of different atoms so their simulated density will be proportional to their mass and as a third option you can even use the charge which might not be so useful if you think of cryo electron microscopy so we will take the partial charge of the atoms assigned here but you could think of some applications involving charge densities for example so some extra thing added here for you to be creative in this case another important parameter is how much weight we actually want to give to the reference density you're trying to bias your simulation against so one of the more tricky parameters to set because if you set the force constant too low you will see nothing happening in your simulation because the force field just overrolls pretty much all the forces that come from the density guided simulation if on contrast you set the force constant too high you will see that the simulation will crash quite quickly because atoms receive all the forces pretty much from the density guided simulations and these forces are much much higher than the force field forces that keep the protein searching in a reasonable shape as their chemistry their chemically and what you would like to do is to find the best middle ground here and there you really have to think about the science you want to do you want to think about the refinement you're doing and here I really recommend just be bold try of things methodologically and try to find the best parameters here after some type of scan so all the city guided simulation and whatever you do we input external experimental data into simulations is a matter of balancing different items of information that is balancing how much do you trust your force field versus how much do you trust your experimental data and there's no exception here so this is really I cannot give you a very general recommended guide another parameter to set is just how blurry do you want to simulate the density to be and there we see at the bottom here three different spreading bit used so you could say that maybe your optimum positions are not extremely well defined and you spread them out a lot you see red very blurry density that is simulated you see a middle ground case in green and you see a very peaky case in blue now if you imagine you choose an extremely slow, low spreading bit of something like 0.01 nanometase then you can imagine that you see only very small peaks at the rock sills so very crisp density so this is something you might want to avoid and on the contrary you also want to avoid spreading out things too far because then all the information you input from the kind of structure side of the system is that there's just one large blob of simulated density and you will receive extremely low forces because blob matches almost anything so here again it's a matter of your system on choosing a best parameter however this seems to be quite robust overall as a parameter and the 0.2 seems to be a good choice for the systems we had on hand so far note also here that this has a very large effect on the compute effort you will spend when running these density-guided simulations as the larger the spreading width the larger area has to be considered for of the density for computation so it is hard to say exactly how things behave in a complex scenario but it might be roughly proportional to the power of 3 so you see there is quite a drastic influence here about the spreading width then you can set the parameter I think that is the longest on all of Gromix parameters and that is the density-guided simulation Gaussian transform spreading range and multiples of it and we try to be explicit here and we thought also we can afford to be quite long here in the parameter name that is one parameter you might very seldomly want to change and this parameter determines where your Gaussian is just cut off so as you see at the bottom left here we have a different cut-off length for this parameter just visualized it doesn't make sense to evaluate this kind of spreading Gaussian at a distance extremely far from the atoms and there is virtually no contribution but at the same time also we don't want to cut it off too short on the right hand side diagram you see the behavior in terms of this parameter how much of the density is included in the overall model if you choose a certain cut-off so if we use a bit of 4 which is the default parameter then we see on the blue line that almost all of the density is included note here also that this is slightly different in three dimensions because they very quickly include more and more density in contrast to just one dimension, one dimension of Gaussian and so 4 is a really good compromise for including more than 99. 9% of the density as you see in the red line reflecting the difference here of what is missing if we go to 2 or 3 we miss still quite some substantial amount and if we go to 5 or 6 we just take away too much but it all means just to also play and get feeling for what this parameter does to use simulations Next parameter to set is just the reference density and there is not so much to note here apart from one exception to the rest of the Gromix philosophy which usually is to condense all input data into one dot tpr file and then be ready to run the simulation. Here is the exception from this rule the idea being that the density files can be quite large and you might not want to bundle and run them and philosophically have more of an idea of ok here is an extra input stream to your simulation and that is this type of density file you would open but this has the effect that at grompp time you can just name whatever file you want and you can just write in this thing whatever string you would like and only when you start here and you run we will see and check that this file is actually present which has the advantage that when you set up your system a utpr file you don't have to have the density file present and has the disadvantage of course that you might get a bad surprise when you run your run simulation and your density file is not in the place where it is to be expected so be very a bit of this you can use absolute or relative plan names also depending on what you prefer here so this might be something to be aware if you especially if you use an absolute file name on your local machine create a tpr file move to the cluster and suddenly the density file is not found this might be the reason in this case here density guided simulations are expensive on the first site because spreading atoms on a grid calculating forces grid forces is something that is hard to keep up in speed with the rest of the force field calculations we have extremely long range interactions we have one global object which is kind of the simulated density that has to be communicated all over the nodes so we want to speed up the process without losing accuracy and here the parameter that allows multiple times will help you a lot this is density guided simulation minus nst this is set to one for safety reasons so to say in our default parameters but for most of the cases you can actually use quite a higher parameter up to 100 I would say just from experience and using this parameter and about 100 usually you will see that the simulation compute time of simulation with density guided forces and simulation without is not that drastic so the difference will be lesser the larger this nst number is so something to to reasonably think but what you would like to think about is what type of movements and what type of oscillation periods have the movements I am guiding the density guided simulation another parameter that most often I would say is just set to through is to normalize densities that is to normalize your reference density and normalize your simulated density just because it allows a way to have comparable force constants for similar systems otherwise some part of the similarity measure is affected by the normalization routine and this will mean that you will have to have change in force constant to catch kind of changes in normalization in density however just make sure if you use some kind of density clipping tools or some tools we say okay I don't care about some part of the density I just set it to large negative values then what you will see this normalization is that these large negative values will have the effect that the sum of all the values in the density will just be negative and then you suddenly normalize with negative value which will result in really weird simulations so just one thing to be aware of anything else I think considered this is a very safe option to just keep on truth then as I mentioned in the beginning you will be faced with the challenge to choose a peak force constant and this might be a very hard task to do right and might be quite legit to go back and forth with different types of force constants so we added another option for the density guided simulations and that is some adaptive force scaling and the idea here is that you start out with quite a low force constant so you would like to set the force constant quite low this option and then we carefully notch up the force if we see that the similarity between simulated and reference density does not improve on the other hand if we see that things are just going fine and we improve the fit then we scale down the force constant a little bit but just not as much as we notch it up and what will happen is that we just go steady and steady and steady to higher and higher and higher similarities now if we went to very high similarity already and we still keep notching up the force constant still not happy we will reach at some point some situation where we just apply very high forces to the system just really to make it fit just a little bit better still and this is quite expected so in this case in contrast to most almost all molecular simulation scenarios you really want to drive your system until it crashes just so you can post process and take the frame most happy with the balance between influence from density guided simulation forces and your structure based forces and again as I said this balance you have to decide on your own and have to say okay do I see the best match between density and and still have a good sense for the structure forces not also if you want to get a nice stereochemistry in the end from structures of MD simulations one thing that works for all types of MD simulation setups is to actually cool down your system afterwards freeze it to lower temperatures to make sure that you take out thermal energy and allow the system to relax again again the setting that is also just possible with density guided simulations as it is compatible with really all types of simulations you set up there's one other parameter in here that is a time constant that is default set to quite impatient 4 picoseconds that is how quickly do we scale up the forces if we see that the fit does not improve and here again it's a matter of performance in simulation times if you want a very very gentle careful fit you can choose a very large times here you will observe that the first part of the simulation that pretty much nothing will be happening and then only very carefully forces will go up or you are quite confident that especially when you structure and density are quite well aligned and you just want a final touch up with the density guided simulations you can choose a quite harsh scaling as is done with the default value of 4 picoseconds here I would like to thank you now we should accept all the things necessary to run density guided simulations as I said before it's mostly really just add density and get a feel for the parameters play a little bit maybe make things crash if you need to I'd like to thank some of my funding agencies by excelling that funded pretty much of the implementation work that was done in gromics and also like to thank Eric Lindahl and all the people in the group especially the gromics developers for looking at my code telling me how to do things better and quite some input also from people also from gutting that ran density guided simulations in their own gromics fork and we had some discussions on how to do things in a good way so Karsten Maxim and then the MPI in gutting and with that we go over it to the questions and answers session where I hope to provide you some insights that I couldn't yet provide you with the seminar. Thank you very much for this very interesting overview of some of the new features in gromics and for talking us through that Christian there are a couple of questions if anyone has any questions that they've not wanted to ask yet please feel free to add them by clicking on the questions tab and writing them up and we'll get to your questions. We will try to go through as many questions as we can the first question we have is from Christopher Linde who if you would like to unmute and ask your question you're welcome to otherwise I will therefore read the question I think Christopher wanted to know the examples that you gave were mainly focused on proteins being fitted to densities I assume the methodology is compatible with RNA structures or RNA protein complexes as well? Yes very much so the whole kind of implementation completely independent of what type of system you're looking at so this would work there's one caveat with RNA systems I think they have kind of a cryeum densities with large densities in the phosphorus region so the way you want to simulate your density might affect the results a bit more than usually with just pure protein complexes so in this case things will work there might be some not so smooth sailing issues with the RNA protein interaction overall because RNA might be a bit too little represented in the types of forces it will receive in the code but this is nothing intrinsic to the code this is something intrinsic in the way how to calculate the forward model if you want to be more complex here I think then we might think of some things to follow up in the future to fit even better but yes it will work on the onset Thank you very much the next question we have is from Ahadrat Desikham give me one second to try to unmute Ahadrat Ahadrat you have been unmuted if you would like to ask your question please could you ask it otherwise I will ask it for you Hi Christian that was a wonderful talk thanks so I was wondering what sort of initial structures do we use for these simulations do we start out with random coils or do we start out with some of the defined protein structures out there? okay yeah that's a very good point so in this case we can't do protein folding just with densities because densities themselves don't have enough structural information and to for example take a random coil and just run the simulation and hope that at some point it will fold into a good protein structure that then also fits the density that will in almost all cases not happen so the starting structure should be some type of reasonable homology model because there's no constraints on secondary structures however you will expect to see that you see some changes and some folding happening that makes the protein structure just more close to what is reflected in the density but it's all matter of what is in the density data and then again a matter of finite compute resources because yeah it will just take too long to fold the protein kind of in an empty simulation even given that extra information from the density thank you so much can I have a quick follow up question if time permits yep go ahead so after you've fed the densities if the secondary structures do not confirm closely to an ideal alpha helix or a beta sheet should we do any sort of minimization or confirmation sampling you can do that if you like this depends on how you want to read your data so the type of ideal alpha helix beta sheets and so on the idea I think it's something that is very much rooted in x-ray crystallography we have the idea of you have a very neat fixed structure at the very low temperatures if you run empty simulations you will always see deviations from this because things have a temperature also in nature and biology so if you want you can just run as I mentioned some type of a needing simulation where you lower the temperature, freeze the system for example also then without the density guided simulation forces added or depending on what you like and then you should see things or you will see things become more idealized in terms of steer chemistry but this depends on how you want to report and what kind of properties you're looking at thank you so much cool the next question we have is from Wozzeck Wozzeck I have unmuted your microphone if you'd like to ask your question hello everyone hi Christian I have a question how much of a manual intervention you need to do with the density and the models before the actual fitting so I've been working with Maxim a little bit for a membrane protein fitting and then there was I don't know like five steps we had to do before to kind of trim the map to make it fit to the initial model more or less like align the model with the map and then that was like the optimal thing for the actual simulation so I was wondering if your tool does it similar way or is it more or less trim line sorry I had been muted and could not unmute myself hi Wozzeck thanks a lot for the commentary yes there are some things you don't have to make the density size fit the simulation box anymore so this was one of the things in the previous code you mentioned there is necessary but you will still have to do some manual intervention to make things fit so this is one of the most focused on routes now for a future improvement of the whole code but something I have finished yet and that is that you get automated rotation and translation of a kind of a rigid body fit of your simulated density to the reference density because we noted that this is one of the major points that are tricky in the setup and they still remain in the setup here that you have to have rough alignment at least of your density and your structure and the whole Gromix setup routine and the density adding routines overall make this something that can be quite challenging at times so we try to improve on this but not for this Gromix release for now you can use any type of density kind of size you want but you still want to make sure that things are roughly aligned okay thanks the next question we have is from Christian Marreiter who does not currently have audio tasks questions so I'll ask it on his behalf first he thanks you for the interesting talk and asks how much more expensive our density guided simulations compared to free ones that's a good point I sort of mentioned and this depends largely on the end step parameter I just mentioned so of course if you apply forces every single step they are quite much more expensive than usual simulations so up to even five times as expensive I would say for a typical system but then this drastically reduces and for most applications it's easily reduced by increasing this number of this end step parameter so for typical cases of brain protein refinement for example and a good balance in terms of accuracy we get kind of 10% overhead 20% overhead I would say in contrast to our usual simulation thank you very much the next question we have is from Brian Smith who asks is transform spreading width the half width yes this is sigma so the Gaussian sigma so this should be should be full width half is it full width half maximum thank you very much the next question we have is from Hugo McDermott-Alpeskin Hugo your microphone is unmuted if you'd like to ask your question I'm sorry if I butchered your surname I'm not sure if you can hear me a fantastic talk it was pretty much covered by the previous question but what's the performance overhead specifically does it work on GPUs or are we not there yet yes it works on GPUs in the sense that whatever simulation you run on GPUs and then at the density of fitting forces everything else will be calculated on the GPUs the density guided simulation forces themselves will not be calculated on the GPUs so you might see a high utilization of GPUs for quite some time of your simulation but then every step you apply density guided simulation forces suddenly the CPU will see quite an uptake and then there'll be some heavy lifting for exactly just this part of the forces fantastic thank you very much the next question we have is from Payne Alexander I've unmuted your microphone if you'd like to ask a question yeah sure can you hear me yes great yeah so I've been thinking about how to get a combination of continuously varying cryoEM density maps or taking advantage of the multiple classes that you can get after doing the processing I'm just wondering if you had ideas about something like replica exchange or umbrella sampling or building some kind of mark off model or something where you're moving from density to density have you thought about applying some method like that yes a lot there's lots of ideas out there and I know there's some group that did for example like multi resolution replica exchange scheme for density guided simulations and yeah overall what you can do in gromix is just one density use simulation plus one density for now and then of course what you can do so that you can go from one state to another state if you have two different densities why is just subsequently applying different densities and kind of guide a system first from density A to density B C and so on and get some type of pathway ideas but then yeah there's more sophisticated ideas but there's a whole zoo of that and we just try to see what is the best way to connect things and hook up things to one another we implemented things in quite modular fashion so if you're interested to have a look at the gromix source code you can see that this whole density guided simulation thing is set up to be rewired and set up in whatever imaginable types of ways for other applications excellent great thank you excellent doc the next question we have is from you do sorry one moment I'm having there we go you your microphone is unmuted if you would like to ask your question sorry you sound a bit distance so if it's okay with you I will ask the question in your stead the questions are is there a very crude guess for the first translation on rotation or density guided simulations from the beginning visualize things it's the best guess it's the best guess it's the best guess it's the best guess it's the best guess it's the best guess see if your density is roughly aligned with your model and then if not you have to do that manually and again this goes back to Wojcik's question yes so there's no initial kind of making things match but we know there is something to work on like first priority actually in this type of code in gromix because there is something that has taken quite some human effort of just having to look at things and making things match great thank you very much the next question we have is from Sangyoon Sangyoon I've unmuted your microphone if you would like to ask the question okay can you hear me yes I hear you very well so my question is how do we know that the simulation structure converges enough to that triom map what kind of output should I look at from the simulation does it print it out the bias potential value in any output file yes it does it prints it out in the energy file and this is an extremely good point and sorry for missing that in my talk because there is just the energy file where it can read off temperature and so on now if you run a simulation with density guided simulations there is a new field appearing that is just saying density energy I think it is called where you can read off of the energy input into the system from the density guided simulation and this directly translates into how well density and the structure fit because the energy input into the system is really force constant times similarity so you will see that very directly and the output in the energy file using GMX energy so the unit is kerosol per mole the unit is SI units so kerosol per mole yes okay thanks the next question we have is from Ali Kousai Ali I okay Ali has keeping in themselves muted so I will ask the question in there stead the question is would it be possible to take differences between two densities i.e. one from a protein in an open state and another closed state to derive vectors that may be used to take a homologous protein from a closed to open state yes you can do that if you have two different densities what you can do similar to what I mentioned before is that you can have a density-guided simulation that brings your protein into the open state and then just take that open state run a new density-guided simulation into the closed state and see how the protein moves there you might want to think of a and I think this may be what the question refers to could you just take kind of a different density and drive the system pretty much like a driving along PCA I can back to this between two states just using exactly one density and this is not possible if I am not mistaken but I would have to think that through a little bit more but the other strategy I think it just works as well and is I think just fair enough here thank you very much the next couple of questions are from Floris van Erden Floris I have unmuted your microphone okay thank you yeah thank you very much and thanks for offering me this possibility as a question I was a little bit I couldn't completely get directly your slide about the periodic boundary conditions so you quickly go over it or because you said only the part closest to the density of the protein will actually feel the density if I remember if I got it correctly yes the yeah I can bring up the slide again and so we have periodic boundary conditions and some type of density and there is a mismatch for example here and this is an extreme example but we have a simulation box that is much smaller than the density and here we have the density box in blue that is much larger than the black simulation box and to ask now okay if we look at this realisation here we have multiple copies of proteins that would be exactly in one density box and the question is okay but which of the copies should be forces from the density should it be all the proteins should it just be one protein and how do we decide what type of protein here in this multiple copies scenario would receive the forces and the criterion to decide okay which one of these copies is the ones closest to the centre of the density and this is just indicated in the thick black line you see the part these are the atoms that receive forces from the density so there is only one to one translation so there is never two atoms two periodic images that receive forces from the density but exactly one periodic image always receives forces from the density and that is one closest to the density centre perfect yeah I get it now so if you just have your boxes more or less equal size there should not be a big problem then no exactly so if the blue box is smaller than your simulation box then there is no issue at all and if if they are similar size also the issue is very small but it's still important to keep in mind this type of behaviour because there is a bunch of densities out there especially from cram that have lots of emptiness around them often you don't even need to clip them to run the densicated simulations you can use them as is but then you have to be aware how the periodic boundary conditions are treated so you might even end up in such a scenario where you have these small black boxes and a large density then placement is virtually everything okay yeah thanks for the help and can I ask the second question and I'm just amazed how well protein can get into such a density and I was wondering does it often get stuck in a local minima this can happen yes so it's not a single push button solution even though we would really love to have that so yeah there's some things that help looking at the different spreading parameters there's some adaptive protocol for example Maxim Igayev had developed as published what helps is using different types of potentials to minimize the kind of ruggedness of the energy landscape for the fitting process but these are all different types of considerations none of them in itself guarantees that you'll never get stuck in the local minima but it makes it easier yeah one thing that might happen also for you runs is to run two simulations three, four density-guided simulations to see which one has the best result so the whole process is stochastic and there will be a local minima that have to be overcome so it's stochastic so I don't even need to have to set a different seed in my normal force field I don't know how you say it this is the normal part of the simulation just to make sure there's no misunderstandings here the forces are deterministic in the way they're calculated but of course we have always numerical noise in our MD simulations that make them quickly diverge on different paths and then the stochasticness was referring to the normal stochasticness in any type of MD simulation and that's exactly what you also see here so if you do run these simulations with Gromax but if you use four different simulations even with the perfectly same studying conditions they will diverge quickly and then you will see different types of a behavior that are on average the same and we know what to expect on average but not for each and every single simulation Perfect, thanks for clarifying and I really like the presentation it was very useful, thanks a lot Thank you The next question we have is a follow-up question from Rajat Rajat, your microphone is unmuted if you'd like to ask your question Hi, Christian Quick question again With the magnitude of forces that you use in density fitting, I was wondering if the choice of force field was immaterial or does it make a difference? Yeah, that's a very good question You can imagine that if you use extremely high forces just from the density there's very much overruled the force field in large parts and I think what happens if you have lots of information from a density is that the choice of the force field makes a bit less of a difference but it still does make a difference so you want to be careful but to say you see more geometric considerations taken into account and less of the kind of stereochemistry considerations depending of course on how large you choose the influence of the reference density to be Do you have favorites among force fields? No, this very much depends on exactly what you want to assimilate so there's a clear answer you cannot have a favorite here because there's a system dependent so for almost all systems there's a good choice but if you can if you have multiple variants then use two force fields as we see in a different case for free energy calculations for example we work with that which is that you're always better if you look at results from different force fields and then at the very end after having done all the simulations look at the meta-analysis of things and I think this might apply here as well Wonderful Thanks so much Thank you The next question we have is from Javier and Javier unfortunately does not have audio so I'll ask the question for him Javier thanks you for the talk and asks do you think we can study protein conformational changes with this kind of tool building PMFs for protein transition for instance Yes, most definitely you can do that because you know exactly what type of energy you put into the system from the density-guided simulation process so you can use all types of analysis methods that are available for other things as methods as well and kind of go a bit into the direction of reading density-guided simulations as some type of using a collective variable for some reasons I would guess if you have more information on hand and different types of information like contacts for example or even kind of target confirmation whatsoever these things might be better suited because they have kind of more information so density information is often ambiguous in a way that we just have a blob and there's different ways to make a protein fit a blob of course depending on the resolution but there's a more clear and direct translation often between structural information or a dihedral angle statistics whatsoever then is in the densities so yes definitely possible but think about if this is really the most efficient way to do things Thank you very much and our last question for today is from Maxim Maxim your microphone is unmuted if you would like to ask your question I will ask the question on Maxim's behalf Maxim asks have you done any performance benchmarking and if so how does the fitting part of the code slow down the simulation compared to vanilla MD runs Yes we have done some benchmarking and this is extremely much depends on the input parameters and the manual that layout like how much influence you will see so there's quite a complex scaling behavior but it slows down the MD run quite a lot if you use n step equals 1 the moment you go up to larger values then much less so but this is kind of a trivial part of the equation then impact comes from how large was spreading with you use how many you take into account and how many atoms are being spread on the grid so it's mostly atoms talking to grid points issue here but the best I can say is yeah it's extremely variable and dependent and it can be up to like that it slows you down 5 fold if you use n step equals 1 to having just a small overhead in the simulation and then the reference manual will just have one formula where we estimate the impact that's all I can say now I think Julian is thank you very much thank you very much Christian thank you for everybody attending now we are closing this section and I hope Christian could you show the next slide so I would like to announce the new the following the seminar that will come after this one so we will have in the 12th of May there will be an investment that will present the clustering free energy landscape from molecular dynamic simulation and then at the end of May exactly the 28th of May we will have Vrindavanet and Benjamin Web that will be presenting PDB DEV a prototype system for achieving integrative structure so please if you are interested please enroll in these two webinars that are already on BioXcel webpage and thank you again for your attention see you next time I hope bye bye